A fan modeling method and system based on a hybrid neural network model

Abstract

The application relates to the technical field of fan modeling, and discloses a fan modeling method and system based on a hybrid neural network model. The fan modeling method based on the hybrid neural network model comprises the following steps: step 1, data acquisition and processing: collecting historical operation data of a wind turbine generator and pre-processing the data to construct a data set; and step 2, model construction: constructing a CNN-LSTM hybrid neural network model, wherein the model comprises a convolutional neural network (CNN) layer for extracting data space coupling features and a long short-term memory (LSTM) layer for capturing long-time time sequence dependence. The application effectively solves the problems of low precision and poor adaptability of existing fan modeling methods by constructing the CNN-LSTM hybrid neural network model and combining a particle swarm optimization algorithm to perform hyperparameter optimization, and has significant technical progress and practical value.

Description

Technical Field

[0001] This invention relates to the field of wind turbine modeling technology, and in particular to a wind turbine modeling method and system based on a hybrid neural network model. Background Technology

[0002] Driven by the global energy transition and the "dual carbon" goal, wind turbines are rapidly developing towards larger scale and greater intelligence. The capacity of individual units and the complexity of operating conditions are significantly increasing, placing higher demands on the modeling accuracy, dynamic response speed, and adaptability of wind turbines. Accurate wind turbine modeling is fundamental to realizing core functions such as pitch control, yaw adjustment, and fault diagnosis, directly affecting the unit's power generation efficiency, operational stability, and service life.

[0003] Traditional wind turbine modeling relies heavily on physical mechanism analysis, requiring the simplification of numerous nonlinear coupling relationships and operating condition assumptions, making it difficult to adapt to complex dynamic scenarios. To overcome the bottleneck of mechanism modeling, data-driven modeling methods have gained widespread attention due to their advantage of building system models solely based on data. However, the following technical problems still exist in existing technologies:

[0004] Patent document CN120068768A discloses a modeling method for doubly-fed induction generator (DFIG) wind turbines based on a fully electromagnetic transient simulation platform. This method achieves accurate simulation of the DFIG by establishing models of the wind turbine, phase-locked loop (PLL), shaft system, and generator. However, this approach suffers from computational efficiency bottlenecks. The fully electromagnetic transient model involves massive high-frequency switching processes and microsecond-level time step calculations, making it difficult to meet the timeliness requirements of real-time simulation and online applications for large-scale wind farms. Furthermore, this approach lacks an online adaptive correction mechanism for physical parameters, leading to systematic errors in the model over time.

[0005] Patent document CN121072357A discloses a method and system for modeling and optimizing wind turbine mechanisms based on neural networks. It establishes a state-space model to ensure physical interpretability and then uses a neural network to correct inaccurate parameters. However, this approach carries the risk of predictive failure in practical applications. When actual operating conditions exceed the distribution range of the training data, the neural network is highly susceptible to misjudging unknown physical dynamics as noise or overfitting. Furthermore, the mechanism model and the neural network are serially coupled; if the basic model has topological deviations, simply fine-tuning the neural network parameters cannot correct structural errors.

[0006] Patent document CN118504613A discloses an improved CNN-LSTM algorithm for short-term power load forecasting, employing an improved particle swarm optimization algorithm to optimize model hyperparameters. However, when applied to the field of power load forecasting, this approach differs fundamentally from wind turbine modeling in its input data, feature extraction methods, and prediction objectives, making it difficult to directly transfer to wind turbine dynamic characteristic modeling. Furthermore, the stochastic inertia weights and adaptive learning factor settings of its improved particle swarm optimization algorithm are completely different from those of this invention.

[0007] In summary, existing wind turbine modeling technologies all suffer from common technical shortcomings: related prediction strategies are limited by the expressive power of shallow machine learning models, easily falling into the trap of static physical constraints, and are constrained by the fragmentation problem of feature fusion mechanisms. Specifically, they struggle to simultaneously consider the time-varying nature of physical mechanisms, the fidelity of raw data under extreme operating conditions, and the deep dynamic coupling relationships between multi-source features. These shortcomings make it difficult for prediction systems to guarantee good tracking accuracy and operational reliability under complex and variable marine wind conditions. Summary of the Invention

[0008] To address the problems in the background technology, this invention proposes a wind turbine modeling method and system based on a hybrid neural network model.

[0009] This invention is achieved using the following technical solution: a wind turbine modeling method based on a hybrid neural network model, comprising the following steps:

[0010] Step 1: Data Acquisition and Processing: Collect historical operating data of the wind turbine generators, preprocess the data, and construct a dataset;

[0011] Step 2, Model Construction: Construct a CNN-LSTM hybrid neural network model, which includes a Convolutional Neural Network (CNN) layer for extracting spatial coupling features of data and a Long Short-Term Memory (LSTM) layer for capturing long-term temporal dependencies of data;

[0012] Step 3, Model Optimization: The particle swarm optimization algorithm is used to globally optimize the hyperparameters of the CNN-LSTM hybrid neural network model to obtain the optimal hyperparameter combination; and,

[0013] Step 4, Model Training and Output: The CNN-LSTM hybrid neural network model is trained using the dataset and the optimal hyperparameter combination to obtain a wind turbine prediction model for predicting the state of wind turbine units.

[0014] Furthermore, in step 1, the collected historical operating data includes wind speed, pitch angle position information, active power, and generator speed; and the collected historical operating data comes from the SCADA system or digital twin platform of the wind turbine, with a data sampling interval of 2 seconds.

[0015] Furthermore, in step 1: missing value correction is performed on the collected data, and the corrected data is normalized to map the data to the interval [0, 1]; the normalization process uses the min-max normalization method, and its formula is:

[0016]

[0017] Furthermore, the structure of the CNN-LSTM hybrid neural network model in step 2 includes:

[0018] The input layer is used to receive preprocessed data;

[0019] A CNN layer contains at least one convolutional layer and one pooling layer, used to extract spatial coupling features from the input data;

[0020] An LSTM layer, containing at least one LSTM layer, is used to learn temporal dependencies in features extracted by CNN layers.

[0021] The output layer is used to output the prediction results.

[0022] Furthermore, in the CNN layer, the convolutional layer uses one-dimensional convolution for local perception and feature extraction, and the pooling layer uses max pooling to reduce the dimensionality of the features output by the convolutional layer.

[0023] Furthermore, the LSTM unit in the LSTM layer includes a forget gate, an input gate, and an output gate;

[0024] The forget gate is used to selectively discard information from the previous cell state.

[0025] The input gate is used to determine the new information in the current input that needs to be saved to the cell state;

[0026] The output gate is used to determine the hidden layer output at the current moment based on the updated cell state.

[0027] Furthermore, in the model optimization of step 3, the hyperparameters optimized by the particle swarm optimization algorithm include: the size of the convolutional neural network kernel, the number of convolutional kernels, the number of neurons in the hidden layer of the long short-term memory network, the learning rate, and the regularization coefficient, with an optimization dimension of 5; and the optimization range of the particle swarm optimization algorithm is set to: convolutional kernel size [2, 7], number of convolutional kernels [8, 64], number of neurons in the hidden layer [10, 200], learning rate [0.001, 0.05], and regularization coefficient [0.0001, 0.01].

[0028] Furthermore, in the model training and output of step 4, the wind turbine prediction model takes the current wind speed and pitch command data, as well as historical power and speed data, as inputs, and outputs the predicted values ​​of active power and generator speed at the current moment.

[0029] Furthermore, in step 4, the model training and output step, the predicted values ​​output by the model are denormalized to restore their physical meaning; the formula for the denormalization process is:

[0030]

[0031] This invention also proposes a wind turbine modeling system based on a hybrid neural network model, comprising:

[0032] The data acquisition and processing module is used to collect historical operating data of wind turbine units, preprocess the data, and construct a dataset.

[0033] The model building module is used to build a CNN-LSTM hybrid neural network model, which includes a convolutional neural network (CNN) layer for extracting spatial coupling features of data and a long short-term memory (LSTM) layer for capturing long-term temporal dependencies of data.

[0034] The model optimization module is used to globally optimize the hyperparameters of the CNN-LSTM hybrid neural network model using the particle swarm optimization algorithm to obtain the optimal hyperparameter combination; and,

[0035] The model training and output module is used to train the CNN-LSTM hybrid neural network model using the dataset and the optimal hyperparameter combination to obtain a wind turbine prediction model for predicting the state of wind turbine units.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] This invention employs the Particle Swarm Optimization (PSO) algorithm, leveraging its randomness and swarm cooperation mechanism. This allows particles to navigate the search space influenced by both their historical best performance and the global optimum, increasing their chances of escaping local traps and finding better parameter combinations globally. Experimental data demonstrates that PSO optimization improves the model's active power prediction accuracy by 73%.

[0038] This invention combines convolutional neural networks (CNNs) and long short-term memory (LSTM) networks. The CNN-LSTM model constructs feature maps from the raw wind power data according to time windows, using these maps as input. Leveraging the efficient feature extraction capabilities of CNNs, more useful information can be extracted. Constructing the extracted vectors into a time-series format and using them as input to the LSTM network better handles long-term data series, reduces information loss, and better reflects the temporal and nonlinear relationships of wind power data. Experimental results show that the R² of active power prediction reaches 0.9733, and the predicted values ​​have a high degree of fit with the actual values.

[0039] This invention captures the long-term temporal dependence of wind speed and rotational speed through a long short-term memory network, and extracts the spatial coupling features of pitch and yaw through a convolutional neural network. It achieves efficient feature fusion of multi-dimensional data, solves the problem of poor adaptability of single neural network algorithms, and can effectively meet the modeling needs of wind turbines under complex and variable operating conditions.

[0040] The method of this invention is data-driven and does not rely on complex physical mechanism analysis. It can be directly applied to the SCADA system or digital twin platform of wind turbines, providing high-precision modeling tools for scenarios such as digital twin modeling of wind turbines, verification of advanced control strategies, and algorithm design.

[0041] In summary, this invention effectively solves the problems of low accuracy and poor adaptability of existing wind turbine modeling methods by constructing a CNN-LSTM hybrid neural network model and combining it with particle swarm optimization algorithm for hyperparameter optimization, and has significant technological progress and practical value. Attached Figure Description

[0042] Figure 1 This is a flowchart of the wind turbine modeling method based on a hybrid neural network model proposed in this invention;

[0043] Figure 2 This is a diagram of the CNN-LSTM model structure in an embodiment of the present invention;

[0044] Figure 3 This is a structural diagram of an LSTM cell according to an embodiment of the present invention;

[0045] Figure 4 This is a flowchart of the PSO algorithm in an embodiment of the present invention;

[0046] Figure 5 This is a flowchart of the PSO-CNN-LSTM algorithm in an embodiment of the present invention;

[0047] Figure 6 This is a graph showing the variation of the loss function during the training process of the CNN-LSTM model in this embodiment of the invention.

[0048] Figure 7This is a comparison chart of the active power output of CNN-LSTM and rotational speed in an embodiment of the present invention;

[0049] Figure 8 This is a comparison chart of active power prediction before and after optimization of the particle swarm optimization algorithm in an embodiment of the present invention. Detailed Implementation

[0050] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0051] Example:

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0053] It is understood that although a logical order is shown in the flowchart, in some cases the steps shown or described may be performed in a different order than that shown in the flowchart. The terms "first," "second," etc., in the specification, claims, or the foregoing drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0054] Reference Figures 1-8 To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in further detail below with reference to the accompanying drawings and specific embodiments. Those skilled in the art should understand that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit its scope.

[0055] Example 1

[0056] This embodiment provides a wind turbine modeling method based on a hybrid neural network model. This method integrates a hybrid model of convolutional neural network (CNN) and long short-term memory network (LSTM) and uses particle swarm optimization algorithm to optimize hyperparameters, thereby solving the technical problems of insufficient accuracy and difficulty in adapting to complex and variable operating conditions in existing wind turbine modeling methods.

[0057] like Figure 1 As shown, the method in this embodiment includes the following steps:

[0058] Step 1: Data acquisition and processing. Collect historical operating data of wind turbines, perform preprocessing, and construct a dataset.

[0059] Step 2: Model building, constructing a CNN-LSTM hybrid neural network model;

[0060] Step 3: Model optimization, using particle swarm optimization algorithm to globally optimize the model's hyperparameters;

[0061] Step 4: Model training and output. Use the dataset and optimal hyperparameters to train the model and obtain the final wind turbine prediction model.

[0062] The following is a detailed explanation of each step.

[0063] 1. Data acquisition and processing steps.

[0064] In this embodiment, the dataset uses wind turbine data from an offshore wind farm in 2025 as the research object. The data sampling time is 8 hours, and the dataset contains 14,000 sets. The data comes from the SCADA system (Supervisory Control and Data Acquisition) deployed locally on the wind turbine or the digital twin platform deployed in the laboratory. Through real-time synchronous acquisition, operational data including historical wind speed, pitch angle position information, active power, and generator speed are obtained, with a sampling interval of 2 seconds.

[0065] Wind power generation and generator speed data, being time series, exhibit certain periodicity and trends. To improve the prediction accuracy and stability of the model, preprocessing of the raw data is necessary before model training. This embodiment's preprocessing includes missing value correction and data normalization.

[0066] First, the collected raw data is examined, and missing values ​​caused by sensor failure or communication interruption are corrected, for example, by using linear interpolation or neighbor-value filling.

[0067] Secondly, to eliminate the impact of different variables on model training due to their different units and to improve the model's computational convergence speed, this invention employs a min-max normalization method to map the data to the [0, 1] interval. The normalization formula is as follows:

[0068]

[0069] in, This is the original data. and These are the minimum and maximum values ​​in the feature data, respectively. This is the normalized data. Through the above processing, a dataset for subsequent model training and testing was constructed.

[0070] 2. Model Building Steps

[0071] The CNN-LSTM hybrid neural network model structure constructed in this embodiment is as follows: Figure 2As shown, this model combines the advantages of CNN in extracting spatially coupled features with the ability of LSTM to process long-term sequences. The model specifically includes the following four-layer structure:

[0072] (1) Input layer: used to receive preprocessed data. In this embodiment, the input of the model is historical data within a specific time window.

[0073] (2) CNN Layer: Used to extract deep spatial coupling features from the input data. In this embodiment, the CNN layer consists of one convolutional layer and one pooling layer. The convolutional layer uses one-dimensional convolution, employing multiple convolutional kernels for local perception and feature extraction. The more kernels, the more abstract the extracted features. The mechanism of local perception and weight sharing effectively reduces model parameters. Its output can be expressed as:

[0074]

[0075] in, For input data, For convolution kernel weights, For bias, For activation function, This is the feature map output by the convolutional layer.

[0076] A pooling layer follows the convolutional layer; in this embodiment, max pooling is preferred. The pooling layer downsamples the feature map output by the convolutional layer, compressing the number of parameters, avoiding overfitting, and enhancing the model's fault tolerance. The output of the pooling layer can be expressed as:

[0077]

[0078] in, This represents the max pooling function.

[0079] (3) LSTM layer: Used to learn the temporal dependencies in the feature vectors extracted by the CNN layers. The LSTM network, through its unique gating structure, can effectively learn long-short-term dependency information. The structure of an LSTM unit is as follows: Figure 3 As shown, its core update formula is as follows:

[0080] Forget Gate: Determines the state of the cell from the previous time step. Which information is discarded?

[0081]

[0082] Input gate: determines the input at the current moment. Which new information will be saved to the cell state?

[0083]

[0084]

[0085] Cell state update: update the cell state from the previous time step. Updated to the current time .

[0086]

[0087] Output gate: Determines the hidden layer output at the current time step based on the updated cell state. .

[0088]

[0089]

[0090] in, It is the sigmoid activation function. The hyperbolic tangent activation function is used. and These are the corresponding weight matrices and bias terms, respectively. This represents the Hadamard product.

[0091] (4) Output layer: The final features extracted by the LSTM layer are mapped to the output space to obtain the final prediction result.

[0092] 3. Model optimization steps.

[0093] In this embodiment, the performance of the CNN-LSTM model is highly dependent on the settings of its hyperparameters. Manually tuning these hyperparameters is not only inefficient but also prone to getting trapped in local optima. Therefore, this invention introduces a particle swarm optimization algorithm to globally optimize the model's key hyperparameters.

[0094] In this embodiment, the hyperparameters to be optimized include five dimensions: the kernel size of the convolutional neural network, the number of convolutional kernels, the number of neurons in the LSTM hidden layer, the learning rate, and the regularization coefficient. The particle swarm optimization algorithm flow is as follows: Figure 4 As shown, the process of optimizing the CNN-LSTM model within the PSO algorithm framework is as follows: Figure 5 As shown.

[0095] The core of the particle swarm optimization algorithm is to simulate the foraging behavior of bird flocks, finding the optimal solution through cooperation and information sharing among individuals. Let the nth... The position of each particle is For a set of hyperparameters to be optimized, the flight speed is... Particles track an individual's historical best position. and the group's historical best position To update its speed and position:

[0096]

[0097]

[0098] in, For inertial weights, and As a learning factor, and It is a random number in the interval [0, 1].

[0099] In this embodiment, the parameters of the PSO algorithm are set as follows: particle swarm size is 10, maximum number of iterations is 15, and inertia weight is... The optimization range was set empirically as follows: kernel size [2, 7], number of kernels [8, 64], number of hidden layer neurons [10, 200], learning rate [0.001, 0.05], and regularization coefficient [0.0001, 0.01]. By iteratively calculating the fitness value (i.e., the prediction error of the model on the validation set) for each particle, the optimal combination of hyperparameters that maximizes model performance was finally found.

[0100] 4. Model training and output steps.

[0101] In this embodiment, the CNN-LSTM model is reconstructed using the optimal hyperparameter combination obtained in step 3, and trained using the dataset constructed in step 1.

[0102] In this embodiment, during model training, the dataset is divided into training and test sets at a ratio of 70% and 30%, respectively. The model takes current wind speed and pitch command data, as well as historical power and speed data, as input, and outputs predicted active power and generator speed for the current moment. To prevent overfitting, an early stopping strategy is used during training; training stops when the loss function value on the validation set no longer decreases over several consecutive epochs.

[0103] During training, the loss function of the model changes as follows: Figure 6 As shown in the figure, both the training loss and the root mean square error (RMSE) of the model show a decreasing trend and tend to stabilize, indicating that the model is well trained.

[0104] Since the model outputs normalized predictions, to give them practical physical meaning, the output needs to be denormalized to restore them to the original data's dimensions. The denormalization formula is:

[0105]

[0106] in, The normalized predicted value output by the model. and These are the maximum and minimum values ​​of the corresponding features in the training data.

[0107] 5. Simulation Analysis and Effect Verification

[0108] To verify the effectiveness of the method of this invention, the trained CNN-LSTM model was tested. The predicted values ​​of active power and generator speed were compared with the actual values, as shown in the figure. Figure 7 As shown in the figure, the blue curve represents the actual value, and the red dashed line represents the model's predicted value. The figure shows that the predicted value fits the actual value well.

[0109] This embodiment uses root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (COP). This is used as an evaluation indicator. Its calculation formula is as follows:

[0110]

[0111]

[0112]

[0113] in, For sample size, For the true value, For predicted values, This represents the average of the true values. The smaller the RMSE and MAPE values, the higher the prediction accuracy. The closer the value is to 1, the better the model fits.

[0114] Calculations show that the predicted active power... The value reached 0.9733, indicating that the model has high prediction accuracy.

[0115] Furthermore, to verify the effectiveness of the Particle Swarm Optimization (PSO) algorithm, this embodiment compares the prediction results of the unoptimized CNN-LSTM model with those of the PSO-CNN-LSTM model optimized by PSO. The active power prediction accuracy is compared, for example... Figure 8 As shown in the figure. The results indicate that after PSO optimization, the model's prediction accuracy for active power was improved by 73%.

[0116] The network parameters obtained before and after optimization are compared in Table 1, and the performance indicators are compared in Table 2.

[0117] Table 1. Network parameters of CNN-LSTM and PSO-CNN-LSTM models Model kernel size Number of convolution kernels Number of hidden layer neurons Learning rate Regularization coefficient CNN-LSTM 3 32 50 0.01 0.001 PSO-CNN-LSTM 5 64 200 0.01516 0.0001

[0118] Table 2 Performance metrics of CNN-LSTM and PSO-CNN-LSTM models Model Active power RMSE Generator speed RMSE Active power R2 CNN-LSTM 153.8880 6.6814 0.9512 PSO-CNN-LSTM 41.5552 2.7730 0.9753

[0119] As can be seen from Tables 1 and 2, after PSO optimization, all performance indicators of the model were significantly improved, and RMSE decreased substantially. Further improvements were made. This demonstrates that the PSO algorithm successfully found a better combination of hyperparameters for the model, effectively improving the model's prediction accuracy and stability.

[0120] In summary, this invention effectively solves the problems of low accuracy and poor adaptability of existing wind turbine modeling methods by constructing a CNN-LSTM hybrid neural network model and combining it with particle swarm optimization algorithm for hyperparameter optimization. It provides a high-precision modeling method for applications such as digital twins of wind turbines and verification of advanced control strategies.

[0121] Example 2

[0122] This embodiment provides a wind turbine modeling system based on a hybrid neural network model, which is used to implement the method described in Embodiment 1. The system includes:

[0123] The data acquisition and processing module is used to collect historical operating data of wind turbine units, preprocess the data, and construct a dataset.

[0124] The model building module is used to build a CNN-LSTM hybrid neural network model, which includes a convolutional neural network (CNN) layer for extracting spatial coupling features of data and a long short-term memory (LSTM) layer for capturing long-term temporal dependencies of data.

[0125] The model optimization module is used to globally optimize the hyperparameters of the CNN-LSTM hybrid neural network model using the particle swarm optimization algorithm to obtain the optimal hyperparameter combination; and,

[0126] The model training and output module is used to train the CNN-LSTM hybrid neural network model using the dataset and the optimal hyperparameter combination to obtain a wind turbine prediction model for predicting the state of wind turbine units.

[0127] The specific implementation details of each of the above modules correspond one-to-one with the method steps in Example 1, and will not be repeated here.

[0128] The above description is merely a preferred embodiment of the present invention and does not limit the scope of the patent. Any equivalent structural modifications made based on the inventive concept of the present invention and the description and drawings, or direct / indirect applications in other related technical fields, are included within the scope of patent protection of the present invention.

[0129] The above provides a detailed description of the preferred embodiments of this application. However, this application is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

[0130] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A wind turbine modeling method based on a hybrid neural network model, characterized in that, Includes the following steps: Step 1: Data Acquisition and Processing: Collect historical operating data of the wind turbine generators, preprocess the data, and construct a dataset; Step 2, Model Construction: Construct a CNN-LSTM hybrid neural network model, which includes a Convolutional Neural Network (CNN) layer for extracting spatial coupling features of data and a Long Short-Term Memory (LSTM) layer for capturing long-term temporal dependencies of data; Step 3, Model Optimization: The hyperparameters of the CNN-LSTM hybrid neural network model are globally optimized using the particle swarm optimization algorithm to obtain the optimal hyperparameter combination; Step 4, Model Training and Output: The CNN-LSTM hybrid neural network model is trained using the dataset and the optimal hyperparameter combination to obtain a wind turbine prediction model for predicting the state of wind turbine units.

2. The wind turbine modeling method based on a hybrid neural network model according to claim 1, characterized in that, In step 1, the collected historical operating data includes wind speed, pitch angle position information, active power and generator speed; and the collected historical operating data comes from the SCADA system or digital twin platform of the wind turbine, with a data sampling interval of 2 seconds.

3. The wind turbine modeling method based on a hybrid neural network model according to claim 1, characterized in that, In step 1: missing values ​​are corrected in the collected data, and the corrected data is normalized to map the data to the interval [0, 1]. The normalization process uses the min-max normalization method, and its formula is: 。 4. The wind turbine modeling method based on a hybrid neural network model according to claim 1, characterized in that, The structure of the CNN-LSTM hybrid neural network model in step 2 includes: The input layer is used to receive preprocessed data; A CNN layer contains at least one convolutional layer and one pooling layer, used to extract spatial coupling features from the input data; An LSTM layer, containing at least one LSTM layer, is used to learn temporal dependencies in features extracted by CNN layers. The output layer is used to output the prediction results.

5. The wind turbine modeling method based on a hybrid neural network model according to claim 4, characterized in that, In the CNN layer, the convolutional layer uses one-dimensional convolution for local perception and feature extraction, and the pooling layer uses max pooling to reduce the dimensionality of the features output by the convolutional layer.

6. The wind turbine modeling method based on a hybrid neural network model according to claim 4, characterized in that, The LSTM cells in the LSTM layer include forget gates, input gates, and output gates; The forget gate is used to selectively discard information from the previous cell state. The input gate is used to determine the new information in the current input that needs to be saved to the cell state; The output gate is used to determine the hidden layer output at the current moment based on the updated cell state.

7. The wind turbine modeling method based on a hybrid neural network model according to claim 1, characterized in that, In the model optimization of step 3, the hyperparameters optimized by the particle swarm optimization algorithm include: the size of the convolutional kernel of the convolutional neural network, the number of convolutional kernels, the number of neurons in the hidden layer of the long short-term memory network, the learning rate, and the regularization coefficient, with an optimization dimension of 5; and the optimization range of the particle swarm optimization algorithm is set as: convolutional kernel size [2, 7], number of convolutional kernels [8, 64], number of neurons in the hidden layer [10, 200], learning rate [0.001, 0.05], and regularization coefficient [0.0001, 0.01].

8. The wind turbine modeling method based on a hybrid neural network model according to claim 1, characterized in that, In the model training and output of step 4, the wind turbine prediction model takes the current wind speed and pitch command data, as well as historical power and speed data, as inputs, and outputs the predicted values ​​of active power and generator speed at the current moment.

9. The wind turbine modeling method based on a hybrid neural network model according to claim 1, characterized in that, In step 4, the model training and output step, the predicted values ​​output by the model are denormalized to restore their physical meaning; the formula for the denormalization process is: 。 10. A wind turbine modeling system based on a hybrid neural network model, characterized in that, include: The data acquisition and processing module is used to collect historical operating data of wind turbine units, preprocess the data, and construct a dataset. The model building module is used to build a CNN-LSTM hybrid neural network model, which includes a convolutional neural network (CNN) layer for extracting spatial coupling features of data and a long short-term memory (LSTM) layer for capturing long-term temporal dependencies of data. The model optimization module is used to globally optimize the hyperparameters of the CNN-LSTM hybrid neural network model using the particle swarm optimization algorithm to obtain the optimal hyperparameter combination; and, The model training and output module is used to train the CNN-LSTM hybrid neural network model using the dataset and the optimal hyperparameter combination to obtain a wind turbine prediction model for predicting the state of wind turbine units.

Images